1 //===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file promote memory references to be register references. It promotes
11 // alloca instructions which only have loads and stores as uses. An alloca is
12 // transformed by using dominator frontiers to place PHI nodes, then traversing
13 // the function in depth-first order to rewrite loads and stores as appropriate.
14 // This is just the standard SSA construction algorithm to construct "pruned"
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "mem2reg"
20 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/AliasSetTracker.h"
27 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/Support/CFG.h"
33 #include "llvm/Support/Compiler.h"
37 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
38 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
39 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
41 // Provide DenseMapKeyInfo for all pointers.
44 struct DenseMapKeyInfo<std::pair<BasicBlock*, unsigned> > {
45 static inline std::pair<BasicBlock*, unsigned> getEmptyKey() {
46 return std::make_pair((BasicBlock*)-1, ~0U);
48 static inline std::pair<BasicBlock*, unsigned> getTombstoneKey() {
49 return std::make_pair((BasicBlock*)-2, 0U);
51 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
52 return DenseMapKeyInfo<void*>::getHashValue(Val.first) + Val.second*2;
54 static bool isPod() { return true; }
58 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
59 /// This is true if there are only loads and stores to the alloca.
61 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
62 // FIXME: If the memory unit is of pointer or integer type, we can permit
63 // assignments to subsections of the memory unit.
65 // Only allow direct loads and stores...
66 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
67 UI != UE; ++UI) // Loop over all of the uses of the alloca
68 if (isa<LoadInst>(*UI)) {
70 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
71 if (SI->getOperand(0) == AI)
72 return false; // Don't allow a store OF the AI, only INTO the AI.
74 return false; // Not a load or store.
83 // Data package used by RenamePass()
84 class VISIBILITY_HIDDEN RenamePassData {
86 typedef std::vector<Value *> ValVector;
89 RenamePassData(BasicBlock *B, BasicBlock *P,
90 const ValVector &V) : BB(B), Pred(P), Values(V) {}
95 void swap(RenamePassData &RHS) {
96 std::swap(BB, RHS.BB);
97 std::swap(Pred, RHS.Pred);
98 Values.swap(RHS.Values);
102 struct VISIBILITY_HIDDEN PromoteMem2Reg {
103 /// Allocas - The alloca instructions being promoted.
105 std::vector<AllocaInst*> Allocas;
106 SmallVector<AllocaInst*, 16> &RetryList;
108 DominanceFrontier &DF;
110 /// AST - An AliasSetTracker object to update. If null, don't update it.
112 AliasSetTracker *AST;
114 /// AllocaLookup - Reverse mapping of Allocas.
116 std::map<AllocaInst*, unsigned> AllocaLookup;
118 /// NewPhiNodes - The PhiNodes we're adding.
120 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
122 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
123 /// it corresponds to.
124 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
126 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
127 /// each alloca that is of pointer type, we keep track of what to copyValue
128 /// to the inserted PHI nodes here.
130 std::vector<Value*> PointerAllocaValues;
132 /// Visited - The set of basic blocks the renamer has already visited.
134 SmallPtrSet<BasicBlock*, 16> Visited;
136 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
137 /// non-determinstic behavior.
138 DenseMap<BasicBlock*, unsigned> BBNumbers;
140 /// BBNumPreds - Lazily compute the number of predecessors a block has.
141 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
143 PromoteMem2Reg(const std::vector<AllocaInst*> &A,
144 SmallVector<AllocaInst*, 16> &Retry, DominatorTree &dt,
145 DominanceFrontier &df, AliasSetTracker *ast)
146 : Allocas(A), RetryList(Retry), DT(dt), DF(df), AST(ast) {}
150 /// properlyDominates - Return true if I1 properly dominates I2.
152 bool properlyDominates(Instruction *I1, Instruction *I2) const {
153 if (InvokeInst *II = dyn_cast<InvokeInst>(I1))
154 I1 = II->getNormalDest()->begin();
155 return DT.properlyDominates(I1->getParent(), I2->getParent());
158 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
160 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
161 return DT.dominates(BB1, BB2);
165 void RemoveFromAllocasList(unsigned &AllocaIdx) {
166 Allocas[AllocaIdx] = Allocas.back();
171 unsigned getNumPreds(const BasicBlock *BB) {
172 unsigned &NP = BBNumPreds[BB];
174 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
178 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
181 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info);
183 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
184 SmallPtrSet<PHINode*, 16> &DeadPHINodes);
185 bool PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
186 void PromoteLocallyUsedAllocas(BasicBlock *BB,
187 const std::vector<AllocaInst*> &AIs);
189 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
190 RenamePassData::ValVector &IncVals,
191 std::vector<RenamePassData> &Worklist);
192 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
193 SmallPtrSet<PHINode*, 16> &InsertedPHINodes);
197 std::vector<BasicBlock*> DefiningBlocks;
198 std::vector<BasicBlock*> UsingBlocks;
200 StoreInst *OnlyStore;
201 BasicBlock *OnlyBlock;
202 bool OnlyUsedInOneBlock;
204 Value *AllocaPointerVal;
207 DefiningBlocks.clear();
211 OnlyUsedInOneBlock = true;
212 AllocaPointerVal = 0;
215 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
217 void AnalyzeAlloca(AllocaInst *AI) {
220 // As we scan the uses of the alloca instruction, keep track of stores,
221 // and decide whether all of the loads and stores to the alloca are within
222 // the same basic block.
223 for (Value::use_iterator U = AI->use_begin(), E = AI->use_end();
225 Instruction *User = cast<Instruction>(*U);
226 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
227 // Remember the basic blocks which define new values for the alloca
228 DefiningBlocks.push_back(SI->getParent());
229 AllocaPointerVal = SI->getOperand(0);
232 LoadInst *LI = cast<LoadInst>(User);
233 // Otherwise it must be a load instruction, keep track of variable
235 UsingBlocks.push_back(LI->getParent());
236 AllocaPointerVal = LI;
239 if (OnlyUsedInOneBlock) {
241 OnlyBlock = User->getParent();
242 else if (OnlyBlock != User->getParent())
243 OnlyUsedInOneBlock = false;
249 } // end of anonymous namespace
252 void PromoteMem2Reg::run() {
253 Function &F = *DF.getRoot()->getParent();
255 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are
256 // only used in a single basic block. These instructions can be efficiently
257 // promoted by performing a single linear scan over that one block. Since
258 // individual basic blocks are sometimes large, we group together all allocas
259 // that are live in a single basic block by the basic block they are live in.
260 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas;
262 if (AST) PointerAllocaValues.resize(Allocas.size());
266 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
267 AllocaInst *AI = Allocas[AllocaNum];
269 assert(isAllocaPromotable(AI) &&
270 "Cannot promote non-promotable alloca!");
271 assert(AI->getParent()->getParent() == &F &&
272 "All allocas should be in the same function, which is same as DF!");
274 if (AI->use_empty()) {
275 // If there are no uses of the alloca, just delete it now.
276 if (AST) AST->deleteValue(AI);
277 AI->eraseFromParent();
279 // Remove the alloca from the Allocas list, since it has been processed
280 RemoveFromAllocasList(AllocaNum);
285 // Calculate the set of read and write-locations for each alloca. This is
286 // analogous to finding the 'uses' and 'definitions' of each variable.
287 Info.AnalyzeAlloca(AI);
289 // If there is only a single store to this value, replace any loads of
290 // it that are directly dominated by the definition with the value stored.
291 if (Info.DefiningBlocks.size() == 1) {
292 RewriteSingleStoreAlloca(AI, Info);
294 // Finally, after the scan, check to see if the store is all that is left.
295 if (Info.UsingBlocks.empty()) {
296 // Remove the (now dead) store and alloca.
297 Info.OnlyStore->eraseFromParent();
298 if (AST) AST->deleteValue(AI);
299 AI->eraseFromParent();
301 // The alloca has been processed, move on.
302 RemoveFromAllocasList(AllocaNum);
309 // If the alloca is only read and written in one basic block, just perform a
310 // linear sweep over the block to eliminate it.
311 if (Info.OnlyUsedInOneBlock) {
312 LocallyUsedAllocas[Info.OnlyBlock].push_back(AI);
314 // Remove the alloca from the Allocas list, since it will be processed.
315 RemoveFromAllocasList(AllocaNum);
319 // If we haven't computed a numbering for the BB's in the function, do so
321 if (BBNumbers.empty()) {
323 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
327 // If we have an AST to keep updated, remember some pointer value that is
328 // stored into the alloca.
330 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
332 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
333 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
335 // At this point, we're committed to promoting the alloca using IDF's, and
336 // the standard SSA construction algorithm. Determine which blocks need phi
337 // nodes and see if we can optimize out some work by avoiding insertion of
339 DetermineInsertionPoint(AI, AllocaNum, Info);
342 // Process all allocas which are only used in a single basic block.
343 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I =
344 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){
345 const std::vector<AllocaInst*> &LocAllocas = I->second;
346 assert(!LocAllocas.empty() && "empty alloca list??");
348 // It's common for there to only be one alloca in the list. Handle it
350 if (LocAllocas.size() == 1) {
351 // If we can do the quick promotion pass, do so now.
352 if (PromoteLocallyUsedAlloca(I->first, LocAllocas[0]))
353 RetryList.push_back(LocAllocas[0]); // Failed, retry later.
355 // Locally promote anything possible. Note that if this is unable to
356 // promote a particular alloca, it puts the alloca onto the Allocas vector
357 // for global processing.
358 PromoteLocallyUsedAllocas(I->first, LocAllocas);
363 return; // All of the allocas must have been trivial!
365 // Set the incoming values for the basic block to be null values for all of
366 // the alloca's. We do this in case there is a load of a value that has not
367 // been stored yet. In this case, it will get this null value.
369 RenamePassData::ValVector Values(Allocas.size());
370 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
371 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
373 // Walks all basic blocks in the function performing the SSA rename algorithm
374 // and inserting the phi nodes we marked as necessary
376 std::vector<RenamePassData> RenamePassWorkList;
377 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
378 while (!RenamePassWorkList.empty()) {
380 RPD.swap(RenamePassWorkList.back());
381 RenamePassWorkList.pop_back();
382 // RenamePass may add new worklist entries.
383 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
386 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
389 // Remove the allocas themselves from the function.
390 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
391 Instruction *A = Allocas[i];
393 // If there are any uses of the alloca instructions left, they must be in
394 // sections of dead code that were not processed on the dominance frontier.
395 // Just delete the users now.
398 A->replaceAllUsesWith(UndefValue::get(A->getType()));
399 if (AST) AST->deleteValue(A);
400 A->eraseFromParent();
404 // Loop over all of the PHI nodes and see if there are any that we can get
405 // rid of because they merge all of the same incoming values. This can
406 // happen due to undef values coming into the PHI nodes. This process is
407 // iterative, because eliminating one PHI node can cause others to be removed.
408 bool EliminatedAPHI = true;
409 while (EliminatedAPHI) {
410 EliminatedAPHI = false;
412 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
413 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
414 PHINode *PN = I->second;
416 // If this PHI node merges one value and/or undefs, get the value.
417 if (Value *V = PN->hasConstantValue(true)) {
418 if (!isa<Instruction>(V) ||
419 properlyDominates(cast<Instruction>(V), PN)) {
420 if (AST && isa<PointerType>(PN->getType()))
421 AST->deleteValue(PN);
422 PN->replaceAllUsesWith(V);
423 PN->eraseFromParent();
424 NewPhiNodes.erase(I++);
425 EliminatedAPHI = true;
433 // At this point, the renamer has added entries to PHI nodes for all reachable
434 // code. Unfortunately, there may be unreachable blocks which the renamer
435 // hasn't traversed. If this is the case, the PHI nodes may not
436 // have incoming values for all predecessors. Loop over all PHI nodes we have
437 // created, inserting undef values if they are missing any incoming values.
439 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
440 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
441 // We want to do this once per basic block. As such, only process a block
442 // when we find the PHI that is the first entry in the block.
443 PHINode *SomePHI = I->second;
444 BasicBlock *BB = SomePHI->getParent();
445 if (&BB->front() != SomePHI)
448 // Only do work here if there the PHI nodes are missing incoming values. We
449 // know that all PHI nodes that were inserted in a block will have the same
450 // number of incoming values, so we can just check any of them.
451 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
454 // Get the preds for BB.
455 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
457 // Ok, now we know that all of the PHI nodes are missing entries for some
458 // basic blocks. Start by sorting the incoming predecessors for efficient
460 std::sort(Preds.begin(), Preds.end());
462 // Now we loop through all BB's which have entries in SomePHI and remove
463 // them from the Preds list.
464 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
465 // Do a log(n) search of the Preds list for the entry we want.
466 SmallVector<BasicBlock*, 16>::iterator EntIt =
467 std::lower_bound(Preds.begin(), Preds.end(),
468 SomePHI->getIncomingBlock(i));
469 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
470 "PHI node has entry for a block which is not a predecessor!");
476 // At this point, the blocks left in the preds list must have dummy
477 // entries inserted into every PHI nodes for the block. Update all the phi
478 // nodes in this block that we are inserting (there could be phis before
480 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
481 BasicBlock::iterator BBI = BB->begin();
482 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
483 SomePHI->getNumIncomingValues() == NumBadPreds) {
484 Value *UndefVal = UndefValue::get(SomePHI->getType());
485 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
486 SomePHI->addIncoming(UndefVal, Preds[pred]);
494 /// DetermineInsertionPoint - At this point, we're committed to promoting the
495 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
496 /// which blocks need phi nodes and see if we can optimize out some work by
497 /// avoiding insertion of dead phi nodes.
498 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
500 // Compute the locations where PhiNodes need to be inserted. Look at the
501 // dominance frontier of EACH basic-block we have a write in.
502 unsigned CurrentVersion = 0;
503 SmallPtrSet<PHINode*, 16> InsertedPHINodes;
504 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
505 while (!Info.DefiningBlocks.empty()) {
506 BasicBlock *BB = Info.DefiningBlocks.back();
507 Info.DefiningBlocks.pop_back();
509 // Look up the DF for this write, add it to PhiNodes
510 DominanceFrontier::const_iterator it = DF.find(BB);
511 if (it != DF.end()) {
512 const DominanceFrontier::DomSetType &S = it->second;
514 // In theory we don't need the indirection through the DFBlocks vector.
515 // In practice, the order of calling QueuePhiNode would depend on the
516 // (unspecified) ordering of basic blocks in the dominance frontier,
517 // which would give PHI nodes non-determinstic subscripts. Fix this by
518 // processing blocks in order of the occurance in the function.
519 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
520 PE = S.end(); P != PE; ++P)
521 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
523 // Sort by which the block ordering in the function.
524 std::sort(DFBlocks.begin(), DFBlocks.end());
526 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
527 BasicBlock *BB = DFBlocks[i].second;
528 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
529 Info.DefiningBlocks.push_back(BB);
535 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
536 // of the writes to the variable, scan through the reads of the variable,
537 // marking PHI nodes which are actually necessary as alive (by removing them
538 // from the InsertedPHINodes set). This is not perfect: there may PHI
539 // marked alive because of loads which are dominated by stores, but there
540 // will be no unmarked PHI nodes which are actually used.
542 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i)
543 MarkDominatingPHILive(Info.UsingBlocks[i], AllocaNum, InsertedPHINodes);
544 Info.UsingBlocks.clear();
546 // If there are any PHI nodes which are now known to be dead, remove them!
547 for (SmallPtrSet<PHINode*, 16>::iterator I = InsertedPHINodes.begin(),
548 E = InsertedPHINodes.end(); I != E; ++I) {
550 bool Erased=NewPhiNodes.erase(std::make_pair(PN->getParent(), AllocaNum));
552 assert(Erased && "PHI already removed?");
554 if (AST && isa<PointerType>(PN->getType()))
555 AST->deleteValue(PN);
556 PN->eraseFromParent();
557 PhiToAllocaMap.erase(PN);
562 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
563 /// replace any loads of it that are directly dominated by the definition with
564 /// the value stored.
565 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
567 StoreInst *OnlyStore = Info.OnlyStore;
568 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
570 // Be aware of loads before the store.
571 SmallPtrSet<BasicBlock*, 32> ProcessedBlocks;
572 for (unsigned i = 0, e = Info.UsingBlocks.size(); i != e; ++i) {
573 BasicBlock *UseBlock = Info.UsingBlocks[i];
575 // If we already processed this block, don't reprocess it.
576 if (!ProcessedBlocks.insert(UseBlock)) {
577 Info.UsingBlocks[i] = Info.UsingBlocks.back();
578 Info.UsingBlocks.pop_back();
583 // If the store dominates the block and if we haven't processed it yet,
584 // do so now. We can't handle the case where the store doesn't dominate a
585 // block because there may be a path between the store and the use, but we
586 // may need to insert phi nodes to handle dominance properly.
587 if (!StoringGlobalVal && !dominates(OnlyStore->getParent(), UseBlock))
590 // If the use and store are in the same block, do a quick scan to
591 // verify that there are no uses before the store.
592 if (UseBlock == OnlyStore->getParent()) {
593 BasicBlock::iterator I = UseBlock->begin();
594 for (; &*I != OnlyStore; ++I) { // scan block for store.
595 if (isa<LoadInst>(I) && I->getOperand(0) == AI)
598 if (&*I != OnlyStore)
599 continue; // Do not promote the uses of this in this block.
602 // Otherwise, if this is a different block or if all uses happen
603 // after the store, do a simple linear scan to replace loads with
605 for (BasicBlock::iterator I = UseBlock->begin(), E = UseBlock->end();
607 if (LoadInst *LI = dyn_cast<LoadInst>(I++)) {
608 if (LI->getOperand(0) == AI) {
609 LI->replaceAllUsesWith(OnlyStore->getOperand(0));
610 if (AST && isa<PointerType>(LI->getType()))
611 AST->deleteValue(LI);
612 LI->eraseFromParent();
617 // Finally, remove this block from the UsingBlock set.
618 Info.UsingBlocks[i] = Info.UsingBlocks.back();
619 Info.UsingBlocks.pop_back();
625 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
626 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
627 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
628 // each read of the variable. For each block that reads the variable, this
629 // function is called, which removes used PHI nodes from the DeadPHINodes set.
630 // After all of the reads have been processed, any PHI nodes left in the
631 // DeadPHINodes set are removed.
633 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
634 SmallPtrSet<PHINode*, 16> &DeadPHINodes) {
635 // Scan the immediate dominators of this block looking for a block which has a
636 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
637 DomTreeNode *IDomNode = DT.getNode(BB);
638 for (DomTreeNode *IDom = IDomNode; IDom; IDom = IDom->getIDom()) {
639 BasicBlock *DomBB = IDom->getBlock();
640 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator
641 I = NewPhiNodes.find(std::make_pair(DomBB, AllocaNum));
642 if (I == NewPhiNodes.end()) continue;
644 // Ok, we found an inserted PHI node which dominates this value.
645 PHINode *DominatingPHI = I->second;
647 // Find out if we previously thought it was dead. If so, mark it as being
648 // live by removing it from the DeadPHINodes set.
649 if (!DeadPHINodes.erase(DominatingPHI))
652 // Now that we have marked the PHI node alive, also mark any PHI nodes
653 // which it might use as being alive as well.
654 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
656 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
660 /// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
661 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
662 /// potentially useless PHI nodes by just performing a single linear pass over
663 /// the basic block using the Alloca.
665 /// If we cannot promote this alloca (because it is read before it is written),
666 /// return true. This is necessary in cases where, due to control flow, the
667 /// alloca is potentially undefined on some control flow paths. e.g. code like
668 /// this is potentially correct:
670 /// for (...) { if (c) { A = undef; undef = B; } }
672 /// ... so long as A is not used before undef is set.
674 bool PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
675 assert(!AI->use_empty() && "There are no uses of the alloca!");
677 // Handle degenerate cases quickly.
678 if (AI->hasOneUse()) {
679 Instruction *U = cast<Instruction>(AI->use_back());
680 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
681 // Must be a load of uninitialized value.
682 LI->replaceAllUsesWith(UndefValue::get(AI->getAllocatedType()));
683 if (AST && isa<PointerType>(LI->getType()))
684 AST->deleteValue(LI);
686 // Otherwise it must be a store which is never read.
687 assert(isa<StoreInst>(U));
689 BB->getInstList().erase(U);
691 // Uses of the uninitialized memory location shall get undef.
694 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
695 Instruction *Inst = I++;
696 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
697 if (LI->getOperand(0) == AI) {
698 if (!CurVal) return true; // Could not locally promote!
700 // Loads just returns the "current value"...
701 LI->replaceAllUsesWith(CurVal);
702 if (AST && isa<PointerType>(LI->getType()))
703 AST->deleteValue(LI);
704 BB->getInstList().erase(LI);
706 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
707 if (SI->getOperand(1) == AI) {
708 // Store updates the "current value"...
709 CurVal = SI->getOperand(0);
710 BB->getInstList().erase(SI);
716 // After traversing the basic block, there should be no more uses of the
717 // alloca: remove it now.
718 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
719 if (AST) AST->deleteValue(AI);
720 AI->eraseFromParent();
726 /// PromoteLocallyUsedAllocas - This method is just like
727 /// PromoteLocallyUsedAlloca, except that it processes multiple alloca
728 /// instructions in parallel. This is important in cases where we have large
729 /// basic blocks, as we don't want to rescan the entire basic block for each
730 /// alloca which is locally used in it (which might be a lot).
731 void PromoteMem2Reg::
732 PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) {
733 DenseMap<AllocaInst*, Value*> CurValues;
734 for (unsigned i = 0, e = AIs.size(); i != e; ++i)
735 CurValues[AIs[i]] = 0; // Insert with null value
737 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
738 Instruction *Inst = I++;
739 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
740 // Is this a load of an alloca we are tracking?
741 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) {
742 DenseMap<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
743 if (AIt != CurValues.end()) {
744 // If loading an uninitialized value, allow the inter-block case to
745 // handle it. Due to control flow, this might actually be ok.
746 if (AIt->second == 0) { // Use of locally uninitialized value??
747 RetryList.push_back(AI); // Retry elsewhere.
748 CurValues.erase(AIt); // Stop tracking this here.
749 if (CurValues.empty()) return;
751 // Loads just returns the "current value"...
752 LI->replaceAllUsesWith(AIt->second);
753 if (AST && isa<PointerType>(LI->getType()))
754 AST->deleteValue(LI);
755 BB->getInstList().erase(LI);
759 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
760 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) {
761 DenseMap<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
762 if (AIt != CurValues.end()) {
763 // Store updates the "current value"...
764 AIt->second = SI->getOperand(0);
765 SI->eraseFromParent();
771 // At the end of the block scan, all allocas in CurValues are dead.
772 for (DenseMap<AllocaInst*, Value*>::iterator I = CurValues.begin(),
773 E = CurValues.end(); I != E; ++I) {
774 AllocaInst *AI = I->first;
775 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
776 if (AST) AST->deleteValue(AI);
777 AI->eraseFromParent();
780 NumLocalPromoted += CurValues.size();
785 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
786 // Alloca returns true if there wasn't already a phi-node for that variable
788 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
790 SmallPtrSet<PHINode*, 16> &InsertedPHINodes) {
791 // Look up the basic-block in question.
792 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
794 // If the BB already has a phi node added for the i'th alloca then we're done!
795 if (PN) return false;
797 // Create a PhiNode using the dereferenced type... and add the phi-node to the
799 PN = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
800 Allocas[AllocaNo]->getName() + "." +
801 utostr(Version++), BB->begin());
802 PhiToAllocaMap[PN] = AllocaNo;
803 PN->reserveOperandSpace(getNumPreds(BB));
805 InsertedPHINodes.insert(PN);
807 if (AST && isa<PointerType>(PN->getType()))
808 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
814 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
815 // stores to the allocas which we are promoting. IncomingVals indicates what
816 // value each Alloca contains on exit from the predecessor block Pred.
818 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
819 RenamePassData::ValVector &IncomingVals,
820 std::vector<RenamePassData> &Worklist) {
822 // If we are inserting any phi nodes into this BB, they will already be in the
824 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
825 // Pred may have multiple edges to BB. If so, we want to add N incoming
826 // values to each PHI we are inserting on the first time we see the edge.
827 // Check to see if APN already has incoming values from Pred. This also
828 // prevents us from modifying PHI nodes that are not currently being
830 bool HasPredEntries = false;
831 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
832 if (APN->getIncomingBlock(i) == Pred) {
833 HasPredEntries = true;
838 // If we have PHI nodes to update, compute the number of edges from Pred to
840 if (!HasPredEntries) {
841 TerminatorInst *PredTerm = Pred->getTerminator();
842 unsigned NumEdges = 0;
843 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) {
844 if (PredTerm->getSuccessor(i) == BB)
847 assert(NumEdges && "Must be at least one edge from Pred to BB!");
849 // Add entries for all the phis.
850 BasicBlock::iterator PNI = BB->begin();
852 unsigned AllocaNo = PhiToAllocaMap[APN];
854 // Add N incoming values to the PHI node.
855 for (unsigned i = 0; i != NumEdges; ++i)
856 APN->addIncoming(IncomingVals[AllocaNo], Pred);
858 // The currently active variable for this block is now the PHI.
859 IncomingVals[AllocaNo] = APN;
861 // Get the next phi node.
863 APN = dyn_cast<PHINode>(PNI);
866 // Verify it doesn't already have entries for Pred. If it does, it is
867 // not being inserted by this mem2reg invocation.
868 HasPredEntries = false;
869 for (unsigned i = 0, e = APN->getNumIncomingValues(); i != e; ++i) {
870 if (APN->getIncomingBlock(i) == Pred) {
871 HasPredEntries = true;
875 } while (!HasPredEntries);
879 // Don't revisit blocks.
880 if (!Visited.insert(BB)) return;
882 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
883 Instruction *I = II++; // get the instruction, increment iterator
885 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
886 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
889 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
890 if (AI == AllocaLookup.end()) continue;
892 Value *V = IncomingVals[AI->second];
894 // Anything using the load now uses the current value.
895 LI->replaceAllUsesWith(V);
896 if (AST && isa<PointerType>(LI->getType()))
897 AST->deleteValue(LI);
898 BB->getInstList().erase(LI);
899 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
900 // Delete this instruction and mark the name as the current holder of the
902 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
905 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
906 if (ai == AllocaLookup.end())
909 // what value were we writing?
910 IncomingVals[ai->second] = SI->getOperand(0);
911 BB->getInstList().erase(SI);
915 // 'Recurse' to our successors.
916 TerminatorInst *TI = BB->getTerminator();
917 unsigned NumSuccs = TI->getNumSuccessors();
918 if (NumSuccs == 0) return;
920 // Add all-but-one successor to the worklist.
921 for (unsigned i = 0; i != NumSuccs-1; i++)
922 Worklist.push_back(RenamePassData(TI->getSuccessor(i), BB, IncomingVals));
924 // Handle the last successor without using the worklist. This allows us to
925 // handle unconditional branches directly, for example.
927 BB = TI->getSuccessor(NumSuccs-1);
931 /// PromoteMemToReg - Promote the specified list of alloca instructions into
932 /// scalar registers, inserting PHI nodes as appropriate. This function makes
933 /// use of DominanceFrontier information. This function does not modify the CFG
934 /// of the function at all. All allocas must be from the same function.
936 /// If AST is specified, the specified tracker is updated to reflect changes
939 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
940 DominatorTree &DT, DominanceFrontier &DF,
941 AliasSetTracker *AST) {
942 // If there is nothing to do, bail out...
943 if (Allocas.empty()) return;
945 SmallVector<AllocaInst*, 16> RetryList;
946 PromoteMem2Reg(Allocas, RetryList, DT, DF, AST).run();
948 // PromoteMem2Reg may not have been able to promote all of the allocas in one
949 // pass, run it again if needed.
950 std::vector<AllocaInst*> NewAllocas;
951 while (!RetryList.empty()) {
952 // If we need to retry some allocas, this is due to there being no store
953 // before a read in a local block. To counteract this, insert a store of
954 // undef into the alloca right after the alloca itself.
955 for (unsigned i = 0, e = RetryList.size(); i != e; ++i) {
956 BasicBlock::iterator BBI = RetryList[i];
958 new StoreInst(UndefValue::get(RetryList[i]->getAllocatedType()),
959 RetryList[i], ++BBI);
962 NewAllocas.assign(RetryList.begin(), RetryList.end());
964 PromoteMem2Reg(NewAllocas, RetryList, DT, DF, AST).run();